LMBP神经网络算法的改进

Improvement of LMBP Neural Network Algorithm

针对LMBP神经网络算法的一些不足,如计算量大,收敛速度慢等特点,提出一些改进方法,即将求逆矩阵G-1移到等式左边,用直接分解法求解,极大地减少了计算量,并且采用变步长代替原来的固定步长.通过结合聚合釜的现场数据集进行故障诊断的仿真实验,结果表明所提改进LMBP故障诊断方法有效、可行.

Aiming at a number of deficiencies on LMBP neural network algorithm, such as calculating capacity and slow convergence characteristics, this article made some improvements of methods, that is, the inverse of matrices G-1 moved to the left of the equation, with direct decomposition method for sol- ving, which greatly reduced the amount of calculation, and used variable step size instead of the original fixed step size. Through the combination of field data sets of polymerization reactor for the simulation experiments of fault diagnosis, the results showed that the improved fault diagnosis methods of LMBP were effective and feasible.